Complementary plugin: closing the calibration loop on what claude-mem captures #3883

I've been running claude-mem alongside a small plugin of my own for a while and they solve different halves of the same problem, so I thought it was worth a show-and-tell.

claude-mem captures what happened. What it doesn't do (by design) is decide what a past correction should change about the next review. A reviewer says "every new public sequence parameter needs validation"; that lands in memory; three PRs later the same miss happens because nothing turned the comment into a check.

Scar (https://github.com/kirilklein/scar) is five slash commands and three markdown files in .claude/ that do that turning:

• /gap logs one line whenever a later stage catches what an earlier one should have (CI catches what passed locally, a human catches what the bot didn't). /gaps clusters the log by root cause and proposes the cheapest fix per cluster (lint config > test mapping > review calibration > lesson).
• /retro <pr> reads human PR comments and writes each miss as a concrete pattern into review-calibration.md. /review loads it every run and tags the findings it produces [calibrated].
• /promote keeps lessons.md honest with a confidence ladder (observation → proven pattern → hard rule), moving entries only on evidence.

Real example from a statsmodels PR: one maintainer comment about a tuple kwarg that should have been array_like with validation became a calibration line, and /review now flags any new sequence kwarg stored raw. Six months of my own gap log (19 entries) showed 6 were review misses the bot caught anyway, which got me to reorder my pipeline to open the PR before the local review.

No runtime, no database; it reads the same repo claude-mem observes. If anyone here has a workflow where claude-mem observations feed into review rules, I'd like to hear how you wire it.